feat(kevin_bridge): multi-mention aggregator with capped conviction boost

services/kevin_signal_bridge/aggregator.py

@@ -0,0 +1,97 @@
+"""Multi-mention windowed aggregation.
+
+Reads kevin_stock_mentions since the cursor, groups by symbol within a
+48h trailing window, applies a conviction boost = boost_per_repeat * extras
+capped at max_boost.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timedelta
+from decimal import Decimal
+from typing import Any, Callable
+
+from sqlalchemy import and_, select
+
+from shared.models.meet_kevin import KevinStockMention
+
+
+@dataclass
+class AggregatedMention:
+    """Mention proxy with effective_conviction set after aggregation."""
+
+    id: int
+    symbol: str
+    action: Any
+    conviction: Decimal
+    time_horizon: Any
+    created_at: datetime
+    rationale_quote: str
+    effective_conviction: Decimal
+
+
+class MentionAggregator:
+    def __init__(
+        self,
+        session_factory: Callable[..., Any],
+        window_hours: int = 48,
+        boost_per_repeat: Decimal = Decimal("0.05"),
+        max_boost: Decimal = Decimal("0.20"),
+    ) -> None:
+        self.session_factory = session_factory
+        self.window_hours = window_hours
+        self.boost_per_repeat = boost_per_repeat
+        self.max_boost = max_boost
+
+    async def fetch_pending(self, since_id: int) -> list[AggregatedMention]:
+        async with self.session_factory() as session:
+            unprocessed = (
+                (
+                    await session.execute(
+                        select(KevinStockMention)
+                        .where(KevinStockMention.id > since_id)
+                        .order_by(KevinStockMention.created_at.asc())
+                    )
+                )
+                .scalars()
+                .all()
+            )
+
+            if not unprocessed:
+                return []
+
+            out: list[AggregatedMention] = []
+            for m in unprocessed:
+                window_start = m.created_at - timedelta(hours=self.window_hours)
+                same_symbol_in_window = (
+                    (
+                        await session.execute(
+                            select(KevinStockMention).where(
+                                and_(
+                                    KevinStockMention.symbol == m.symbol,
+                                    KevinStockMention.created_at >= window_start,
+                                    KevinStockMention.created_at <= m.created_at,
+                                )
+                            )
+                        )
+                    )
+                    .scalars()
+                    .all()
+                )
+                extras = max(0, len(same_symbol_in_window) - 1)
+                boost = min(self.max_boost, self.boost_per_repeat * extras)
+                effective = min(Decimal("1.0"), m.conviction + boost)
+                out.append(
+                    AggregatedMention(
+                        id=m.id,
+                        symbol=m.symbol,
+                        action=m.action,
+                        conviction=m.conviction,
+                        time_horizon=m.time_horizon,
+                        created_at=m.created_at,
+                        rationale_quote=m.rationale_quote,
+                        effective_conviction=effective,
+                    )
+                )
+            return out